OBJECTIVES: While the follow-up time on Ipilimumab trials is in excess of 4 years, HTA models often require survival to be extrapolated to 10 years and beyond. However, patient level data on prognostic factors are rarely available; hence extrapolation methods assume a homogeneous study population and are based on statistical considerations only. Such approaches are criticized for disregarding clinical reality and may be biased. In this study a survival extrapolation model that accounted for heterogeneity was developed based on both statistically and clinically relevant considerations. The method was applied on survival data in patients with Metastatic Melanoma. METHODS: Survival data were taken from a randomized controlled clinical trial that compared dacarbazine plus placebo versus dacarbazine plus ipilimumab. Two parametric models were explored to extrapolate survival: a model assuming no heterogeneity in patients and another model that divided patients into three subgroups based on cancer stage observed at baseline and additionally included subpopulations of a priori unobserved long-term survivors. Survival of the subpopulations was extrapolated and summed to obtain survival in the overall population. Subgroup formation was guided by expert opinion of oncologists. The statistical and clinical validity of the models were assessed. RESULTS: Among commonly used distributions (exponential, Weibull, lognormal) the lognormal distribution fitted the survival data best in the no-heterogeneity model whereas Weibull distribution was used for the heterogeneity model. For statistical validity, both models fitted the data reasonably well. However, the no-heterogeneity model underestimated the long tail of the survival curves. The no-heterogeneity model implied decreasing mortality over time while the heterogeneity model implied increasing mortality, which is more clinically relevant. CONCLUSIONS: The no-heterogeneity model fitted the data reasonably well but was not relevant for extrapolation from a clinical perspective. The heterogeneity model captured the long tail of the survival curve best, and provided a statistically and clinically relevant model.